This dissertation presents the practical techniques for structural health monitoring and damage identification of large civil infrastructures using neural networks technique. Neural networks techniques are employed in this study since they are computationally quick to estimate the damage severities for on-line monitoring and assessment and they can utilize various kinds of input data measured during the tests. It consists of two parts: i) development of a health monitoring method for bridges under ordinary traffic loadings using the neural networks technique and ii) enhancement of the damage estimation method for building frames using committee of neural networks.
First, a method for structural health monitoring of bridges is presented including damage assessment based on ambient vibration data caused by the traffic loadings, and it is verified by tests on a bridge model. The procedure consists of the identification of the operational modal properties using the random decrement method, the updating of the baseline finite element model using the inverse modal perturbation technique, and the assessment of the damage locations and severities using the neural networks technique and the committee technique for neural networks. For the practical application of the structural health monitoring based on the vibration induced by operational loads, the sensitivity of the measured vibration data to changes in operating loads should be considered, and structural identification must be carried out with the changes of the modal parameters only due to the damages. In this study, as input to the neural networks for the damage estimation, the ratios of resonant frequencies between before and after the damages and the mode shapes after the damages are used. The frequency ratios have been used instead of the frequencies since the resonant frequencies of the bridge extracted from the vibration data vary depending on the mass of the moving vehicles. An experimental study is carried out on a bridge model with a composite cross-section subjected to vehicle loadings. Vertical accelerations of the bridge deck are measured while vehicles are running. The frequency ratios are found to be effective in reducing the mass effects of the moving vehicles on the operational modal parameters. It has been found that the identified damage locations and severities agree reasonably well with the inflicted damages on the structure.
Secondly, a method for damage estimation of building structures using the committee technique for neural networks is presented for the purpose of structural health monitoring. For practical application, this method aims to monitor the critical members with a limited number of the measurement locations. Multiple neural networks are constructed and each individual neural networks is trained independently with different initial synaptic weights. Then, the estimated damage indices from different neural networks are averaged. Several committee methods are investigated and used to estimate the element-level damage locations and severities for various damage cases with noise-free and noised input data. The non-uniqueness of the solution due to the local minima during the parameter estimation, noise, and limited number of measurements may be resolved by employing the committee technique, which is a statistical approach averaging the damage indices in the functional space. The committee technique is found to be very effective to improve the damage estimation. It has been found that the damage missing and false alarm errors can be reduced significantly, and the accuracy of the estimation can be improved by applying the committee of neural networks. It has also been found that the damages can be successfully identified with the noised input data if the noise injection learning and the committee technique are carried out simultaneously.
No established procedure exists for health monitoring of large structures with damage identification. In this study, a procedure is presented for health monitoring of bridges including damage assessment for which the frequency ratios are used to reduce the mass effects of the moving vehicles on the operational modal properties, and the whole procedure is verified by tests on a bridge model. In addition, the committee technique for neural networks, which has been widely used for pattern recognitions in speech and vision studies, is employed for the structural identification. Effectiveness of the several committee techniques is verified for the damage estimation of a building structure.
본 연구에서는 신경망기법을 이용한 대형구조물의 건전성감시 및 손상추정의 실용화기법에 관하여 연구하였다. 신경망기법은 손상추정에 소요되는 시간이 짧으며, 다양한 형태의 입력자료 활용이 가능하므로 상시감시에 적용되었다. 본 논문은 다음과 같이 크게 두 가지 내용으로 구성되어있다.
첫번째, 손상추정을 수반하는 교량의 건전성감시 방법을 제안하였다. 이 방법은 차량하중에 의한 상시진동기록을 이용하며, 모형교량 실험을 통하여 전 과정이 검증되었다. 차량하중에 의한 상시진동기록으로부터random decrement기법을 사용하여 자유진동신호를 구하며, 이를 이용하여 구조물의 모드특성을 구한다. 손상이 없는 구조물에 대한 유한요소모델을 모드섭동법을 이용하여 개선하며, 신경망기법과 신경망 committee기법을 적용하여 손상위치 및 손상정도를 추정한다. 작용하중에 의한 진동기록을 이용하는 건전성감시 기법의 실용화를 위해서는, 작용하중이 구조물의 모드특성에 미치는 영향을 배제하고 손상에 의한 모드특성의 변화만을 이용해야 한다. 이동 차량의 질량효과로 인한 교량의 공진주파수 변화를 고려하기 위하여, 손상전후 공진주파수 비율을 손상 후 모드형상과 함께 신경망의 입력자료로 이용하여 손상을 추정하였으며, 손상전후 공진주파수 비율 사용의 타당성을 모형교량 실험을 통하여 검증하였다. 제안된 방법을 적용하여 모형교량의 손상위치 및 손상정도를 추정하였으며, 추정된 결과는 실제 손상과 잘 일치하였다.
두번째, 신경망 committee기법을 이용한 빌딩구조물의 손상추정방법을 제안하였다. 즉, committee기법을 도입하여 손상추정의 정확성을 개선하였다. 건전성감시의 실용화를 위하여 제한된 계측자료를 이용하여 주요부재만을 감시하는 방법을 적용하였다. Committee기법에서는 다수의 신경망을 구성하여 각 신경망을 독립적으로 훈련시킨 후, 개별 신경망으로부터 추정된 손상지수들을 평균한다. 이때, 연결강도(synaptic weights)의 초기치를 모두 다른 값으로 설정한다. 여러 가지 committee방법을 제안하였으며, 이를 적용하여 대상 빌딩의 다양한 수치적 손상경우에 대한 손상추정을 수행하였다. 이때, 입력자료에 노이즈가 없는 경우와 있는 경우를 각각 고려하였다. Committee기법은 손상지수을 평균하는 통계적 기법 중 하나로서, 국소최소, 노이즈, 제한된 계측자료 등으로 인해 발생하는 해의 비유일성(non-uniqueness)에 의한 오차를 감소시킬 수 있다. Committee기법을 도입함으로써 damage missing과 false alarm문제를 개선시킬 수 있었고 손상추정의 정확성을 향상시킬 수 있었다. 또한, 입력자료에 노이즈가 포함된 경우 노이즈첨가학습과 committee기법을 동시에 적용함으로써 손상을 성공적으로 추정할 수 있었다.
